Most life on Earth has never been described. In a sample of seawater, most of the DNA belongs to organisms that don't have a name, and the usual tools skip over it.
Below 500 metres, 93.8% of the sequences in public ocean data don't match any named species. We started bioX to work on that part.
Our model learned DNA the way a language model learns text. It read 1.13 million ribosomal RNA gene sequences and learned to predict each next base.
Almost half of that data comes from environmental samples, so it has learned from DNA that reference databases leave out.
Every sequence becomes a position on a map of life: we count how often each of the 256 four-letter words occurs, and related organisms end up close together even when the exact species isn't in any database.
We trained a ten-million-parameter network to do this better. On 40 groups held out of training it placed 42% of reads correctly against 82% for counting, so counting is what we serve. The network is still what scores and writes.
It also shows which parts of a gene tend to stay the same across species, which is where primers bind.
We tried to make it spot sequences stitched together by PCR too; on a fair test it scored 0.53 where 0.50 is guessing, so that one is gone rather than parked.
We think results should be checkable. We test the model on organisms it has never seen, and we show where it gets things wrong.
Every result records which model version produced it, so someone else can reproduce it.
The model is trained on public reference data: SILVA, PR2, and environmental reads from ENA. There is no way to bring your own data yet.
Every model page has a box that runs one sequence and keeps nothing — no account, no storage. Research groups can request API access, and we read every request ourselves.